Operant and classicalconditioning of Drosophila at

the flight simulator

JULIUS-MAXIMILIANS-UNIVERSITÄT WÜRZBURGFAKULTÄT FÜR BIOLOGIE

DIPLOMARBEIT

vorgelegt von

Björn Brembs

Würzburg, August 1996

1

CONTENTS1 INTRODUCTION...............................................................................................................................................2

1.1 LEARNING ...........................................................................................................................................................21.2 CLASSICAL CONDITIONING ..................................................................................................................................21.3 OPERANT CONDITIONING ....................................................................................................................................31.4 HOW DISTINCT ARE CLASSICAL AND OPERANT CONDITIONING? .........................................................................31.5 DROSOPHILA IN THE NEUROSCIENCES.................................................................................................................41.6 DROSOPHILA IN THE FLIGHT SIMULATOR ............................................................................................................4

1.6.1 The Flight Trace..........................................................................................................................................41.6.2 Input/Output Analysis..................................................................................................................................5

1.7 CLASSICAL AND OPERANT CONDITIONING IN DROSOPHILA .................................................................................5

2 MATERIAL AND METHODS..........................................................................................................................5

2.1 THE ANIMALS .....................................................................................................................................................52.2 THE EXPERIMENTAL SETUP ................................................................................................................................62.3 THE EXPERIMENTS..............................................................................................................................................6

2.3.1 The Standard Experiment............................................................................................................................62.3.2 Classical Conditioning................................................................................................................................7

2.4 THE EVALUATION ...............................................................................................................................................72.4.1 The Flight Trace I: Arena Position .............................................................................................................7

2.4.1.1 Avoidance/Learning ...............................................................................................................................................72.4.1.2 Fixation ..................................................................................................................................................................72.4.1.3 Quadrant Changes and Arena Rotation ..................................................................................................................7

2.4.2 The Flight Trace II: Yaw Torque ................................................................................................................72.4.2.1 Spike Detection ......................................................................................................................................................82.4.2.2 Spike Dynamics and -Timing .................................................................................................................................82.4.2.3 Spike Polarity .........................................................................................................................................................8

2.4.3 The Flight Trace III: Combined Evaluation................................................................................................92.4.4 Spike Detection Efficiency...........................................................................................................................9

2.5 STATISTICS..........................................................................................................................................................9

3 RESULTS AND DISCUSSION..........................................................................................................................9

3.1 SPIKE DETECTION ...............................................................................................................................................93.2 SPIKES...............................................................................................................................................................103.3 STEPWISE ARENA ROTATION.............................................................................................................................113.4 MEASUREMENTS AT T1 ......................................................................................................................................123.5 COMPARING THE STANDARD EXPERIMENT AND CLASSICAL CONDITIONING .....................................................12

3.5.1 Avoidance and Learning ...........................................................................................................................133.5.2 Spike Dynamics and -Timing.....................................................................................................................143.5.3 Spike Polarity ............................................................................................................................................163.5.4 Variables Measured Independently of Quadrant Treatment.....................................................................17

3.6 MEASUREMENTS AT T2 ......................................................................................................................................18

4 CONCLUSION..................................................................................................................................................20

4.1 HOW ‘CLASSICAL’ IS THE STANDARD EXPERIMENT? ........................................................................................214.2 CLASSICAL AND OPERANT CONDITIONING: MERELY AN OPERATIONAL DISTINCTION? .....................................214.3 WHAT IS LEARNED DURING CONDITIONING? ....................................................................................................21

5 SUMMARY .......................................................................................................................................................23

6 ACKNOWLEDGEMENTS..............................................................................................................................23

7 ZUSAMMENFASSUNG...................................................................................................................................23

8 REFERENCES..................................................................................................................................................23

APPENDIX ...........................................................................................................................................................26

2

"Under carefully controlled experimental circumstances,an animal will behave as it damned well pleases."

Harvard Law of Animal Behavior

1 IntroductionThe brain’s primary objective is to carry out certainadaptive behaviors. It is fine-tuned by evolution tosafely govern its carrier through life and to achievesuccessful reproduction. There are two alternativesto accomplish this task: by innate behavior pro-grams (e.g. reflexes, stimulus-response chains etc.)adapted by evolution or acquired behavioral traits,adapted by experience.This distinction is not new: in the 18th century, partof the empiricist philosophy of Locke (1689) wasthe assertion that individuals were born with a tab-ula rasa and only experience could establish mind,consciousness and the self. On the continent, Leib-niz envisaged the self as a monad carrying someknowledge of a basic understanding of the world.The discussion as to whether nature or nurture arethe driving force shaping cognitive abilities was fora long time considered to be interminable. Until the1960s this dispute was still very vivid in the be-havioral sciences: in the tradition of the Englishempiricists, Skinner’s school of behaviorism postu-lated general rules for all types of learning, ne-glecting innate differences or predispositions. Lo-renz was one of the protagonists of ethology inEurope, which focused on the inherited aspects ofbehavior. It was Lorenz who ended these antago-nistic views of behavior in showing that there in-deed are innate programs ("fixed action patterns")and predispositions in behavior where only littlelearning occurs. Today, it is largely agreed uponthat nature and nurture are intimately cooperating tobring about adaptive behaviors. Probably only invery few cases ontogenetic programs are not at allsubjected to behavioral plasticity. Conversely, thepossibility of acquiring behavioral traits has to begenetically coded for.

1.1 LearningNevertheless, acquired behavioral traits can in prin-ciple be distinguished from inherited programsoperationally: whereas many innate behaviors aredisplayed even in total deprivation of experience(„Kaspar-Hauser“ experiments), learned behavior isalways absent under such circumstances. Conse-quently, learning can be defined as the process bywhich an organism benefits from experience so thatits future behavior differs from that of a comparableorganism lacking this experience.Typically, studies of learning compare the behaviorof two subjects at two times. At a time t1 the indi-

viduals share the same experience and thus do notdiffer in performing the behavior in question. At alater time (t2), the behavior of the same subjects iscompared again: one of the subjects has in themeantime been exposed to the experience of interest(most commonly the presentation of one or severalstimuli), whereas the other was spared this particu-lar exposure and instead received a control treat-ment. Learning is assessed according to the differ-ence in the behavior of the subjects in t2.By convention, learning is classified operationallyinto three types of stimulus presentation between t1

and t2:1. Presentation of the stimulus alone. → habitua-

tion, sensitization (non-associative learning).2. Presentation of the stimulus in relation to an-

other stimulus. → classical conditioning (asso-ciative learning)

3. Presentation of the stimulus in relation to someof the organisms own behavior. → operant con-ditioning (associative learning)

Since the present work is concerned with a com-parison of classical and operant conditioning, theseare examined more closely.

1.2 Classical ConditioningThe term "classical conditioning" is used here todescribe a type of associative learning in whichthere is no contingency between response and rein-forcer. This situation resembles most closely theoriginal experiment of Pavlov (1927), who traineddogs to associate a tone with a food-reward. In suchexperiments, the subject shows a weak or no re-sponse to a conditioned stimulus (CS, e.g. a tone),but a measurable unconditioned response (UR, e.g.saliva production) to an unconditioned stimulus(US, e.g. food) at t1. In the course of the training,the CS is repeatedly presented together with the US;eventually the subject forms an association betweenthe US and the CS. In a subsequent test-phase (t2),the subject will show the conditioned response (CR,e.g. saliva production) to the CS alone, if such anassociation has been established and memorized.The subject is said to have learned about salientcontingencies in the world. Control subjects usuallyreceive unpaired CS and US presentations or CSand US alone. Such "Pavlovian" conditioning isopposed to instrumental or "operant conditioning",as described below (1.3), where producing a CRcontrols the US presentations.Findings from Kandel and coworkers (Kandel et al.1983; Hawkins et al. 1983; Carew et al. 1983; Ca-

3

rew and Sahley, 1986 and references therein) in-vestigating the cellular and molecular processesunderlying classical conditioning in Aplysia suggestthat the US is ‘replaced’ by the CS during training:simultaneous stimulation of the sensory neuronreceiving the CS (SN1) from the sensory neuronreceiving the US (SN2), facilitates synaptic efficacyof the SN1 presynaptically (Fig. 1). After a fewconditioning trials, stimulation of the SN1 aloneelicits the reflexive behavior - the UR elicitingproperties of the reinforcer have been transferred toSN1.

Fig. 1: Schematic diagram of the classical conditioning modelinferred from evidence in Aplysia. According to the model,activity dependent presynaptic facilitation (ADPF) is induced bythe conjoint firing of the sensory neurons of the CS+ pathwayand the facilitatory interneurons mediating the US. After condi-tioning, firing of SN1 alone will activate the motor neuron. CS+- conditioned stimulus contiguous with the US; CS- - condi-tioned stimulus unpaired with the US; SN1 - sensory neuronreceiving the CS+, SN2 - sensory neuron receiving the US; SN3- sensory neuron receiving the unpaired control stimulus CS-.(Redrawn from Glanzman ,1995)

1.3 Operant ConditioningThe term "operant conditioning" is used here todescribe a type of associative learning in whichthere is a contingency between the presentation ofresponse and reinforcer. This situation resemblesmost closely the classic experiments of Skinner(1938), where he trained rats and pigeons to press alever in order to obtain a food reward ("Skinner-Box"). In such experiments, the subject is often ableto generate a variety of motor-outputs. The experi-menter chooses a suited output (the response R, e.g.pressing a lever) at t1 to pair it with an uncondi-tioned stimulus (US, e.g. a food reward). Often adiscriminative stimulus (SD, e.g. a light) is presentwhen the R-US contingency is true. After a trainingperiod (at t2), the subject will show enhanced R -the conditioned response (CR, e.g. pressing a lever)- even in absence of the US if the R-US associationhas been memorized. The subject is said to havelearned about salient contingencies between its ownbehavior and some part of the world. Such instru-mental or operant conditioning is opposed to Pav-lovian or "classical conditioning", as describedabove (1.2), where producing a response has noeffect on US presentations. This fundamental dif-ference has an important temporal consequence:whereas it is per definitionem of no value for the

subject to make any associations during classicalconditioning - it will receive the US anyway - oper-ant behavior provides the subject with a means tooptimize its situation already during acquisition. Inother words: classical conditioning can be perceivedas passive learning from events in the past, whileoperant conditioning implies learning to behave inthe present and the future.This view is reflected in the work of Wolf and He-isenberg (1991), who have further analyzed theprocess of operant conditioning in Drosophila andpropose a basic model of operant behavior (Fig. 2):

1. Operant behavior requires a goal (desired state).2. In order to achieve the goal, a range of motor

programs is activated (initiating activity).3. Efference copies of the motor programs are

compared to the sensory input referring to thedeviation from the desired state.

4. In case of a significant coincidence, the respec-tive motor program is used to modify the sen-sory input in the direction toward the goal.

Fig. 2: General model of operant behavior. The brain generates alarge variety of motor-outputs and cross-correlates them withone or several sensory inputs. If, for a certain combination, thecorrelation is sufficiently positive, the fly can manipulate thissensory input according to its needs. The long external arrowcoupling 'rotatory control maneuvers' and 'temperature', depictsthe situation of a fly in the Drosophila Flight Simulator. (From:Wolf and Heisenberg, 1991)

Consistent control of a sensory stimulus (i.e. thereinforcer) by a behavior may lead to a more per-manent behavioral change (conditioning).According to Wolf and Heisenberg (1991) operantbehavior is the active choice of one out of severaloutput channels in order to minimize the deviationsof the current situation from a desired situation (1-4). Operant conditioning in these terms is expressedby persisting activation of this channel after thesituation has changed (5). Mutant analyses inDrosophila have shown that these processes canalso be genetically distinguished (Eyding, 1993;Weidtmann, 1993).

1.4 How Distinct are Classical andOperant Conditioning?For a long time classical and operant conditioningwere considered distinct categories of learningrequiring distinct pathways in the brain. Some ob-servations, however, seem to indicate that most

4

learning situations contain operant and classicalcomponents at various degrees.Pavlov’s hungry dog, for example, will show appe-titive behavior towards the bell, if it has seen itwhile it was ringing during classical training even ifit was immobilized at that time. More spectacularly,Chimpanzee males, trained operantly to insert coinsinto a food dispenser will trade the coins to equallytrained females for sex, as they normally do withfood.A key to resolving the dilemma whether operantand classical conditioning can indeed be conceivedas separate entities, might be to compare the differ-ent types of associations the subjects form betweenthe various stimuli, the response and the reinforcer.According to Pavlov (1927), Kandel and his co-workers (Kandel et al. 1983; Hawkins et al. 1983;Carew et al. 1983; Carew and Sahley, 1986) andHammer (1993), stimulus-substitution seems toaccount for the capacity of the CS to elicit the CR inclassical conditioning. The association is assumedto be stimulus-reinforcer in nature. According toSkinner (1938), Mackintosh (1975) and Wolf andHeisenberg (1991), associations in operant condi-tioning are formed between the behavioral output ofthe organism and its stimulus situation. The asso-ciation is assumed to be response-reinforcer innature.

1.5 Drosophila in the Neuro-sciencesThe key insight that all levels of functional organi-zation from genes to behavior tightly interact, con-stitutes the basis for a very successful science: neu-rogenetics. Clearly emphasizing the inherited as-pects of the phenomena studied in neuroscience, theprimary model system for neurogeneticists isDrosophila. The rich repertoire of classical geneticstogether with very efficient molecular techniques,allows one not only to identify and clone new genesbut also to assess their function at the molecular,cellular and systemic level. Exploiting these op-portunities Drosophila offers has furthered to agreat extent our understanding of the molecularmechanisms underlying such complex processes asclassical olfactory conditioning in Drosophila (e.g.Davis and Dauwalder, 1991; Tully et al. 1994;Tully, 1991; Tully, et al. (1990).Besides the genetic level, Drosophila providesseveral advantages for studying learning and mem-ory behaviorally, compared to humans and othermammals: 1) a short lifespan in standardized vialsreduces the inter-individual variance in experientialhistory to a minimum; 2) only minor ethical consid-erations have to be taken into account for experi-mental design; 3) relatively small experimentalsetups; 4) no social or linguistic complications; 5)the possibility to measure a large number of indi-viduals.

The study presented here takes advantage of thesefeatures: Drosophila is used in an experimentalsituation that allows for minute control of the inputthe fly receives and the output it produces.

1.6 Drosophila in the Flight Simula-torIn the flight simulator a single tethered Drosophilafruitfly flies stationarily in an artificial environment.Originally, open loop experiments, in which thefly's behavior has no effect upon its visual stimulussituation, were utilized for detailed examination ofDrosophila’s optomotor behavior (e.g. Heisenbergand Wolf, 1984; Heisenberg and Wolf, 1993). Inthis setup, however, the fly can be enabled to con-trol some aspects of its visual input by coupling itto its motor-output (closed loop). Both the visualinput and the motor-output are monitored on-linethroughout the experiment.Such a flight simulator setup is ideally suited for adetailed comparison of classical and operant condi-tioning, since various contingencies among behav-ioral output, visual input and the reinforcer - eve-rything in exquisite control of the experimenter -can be established. In the present study, the envi-ronment consists of a cylindrical panorama arrangedto center the fly within the cylinder. The motion ofthe environment is limited to the horizontal plane:only the rotational speed of the cylinder can becontrolled by the fly's tendency to turn around itsvertical body axis (yaw torque, see Fig. 3).

1.6.1 The Flight TraceUpon observing a fly in the flight simulator, it isstriking that the fly neither keeps the cylinder im-mobilized nor rotates it continuously: phases offairly straight flight are interrupted by sudden turnsat high angular velocity. Monitoring the fly's yaw

Fig. 3: The two spherical coordinates Ψ and ϑ of a fly’s visualspace and its three degrees of freedom for rotation (yaw, roll,pitch; redrawn from Heisenberg and Wolf, 1984).

5

torque, the turns are due to short pulses of torque(torque spikes, Fig. 4). The fly generates thesetorque spikes by reducing the wing beat amplitudeby about 12° on the side to which it intends to turn(wing hitch; Götz et al. 1979; Götz, 1983). Thesesudden turns („body saccades“) can be observed infree flying Drosophila as well. Based on previouswork from Heisenberg and Wolf (1979), Heisen-berg and Wolf (1984), Mayer et al. (1988) andHeisenberg and Wolf (1993), the present studyassumes that the spikes are the primary behavior bywhich the fly adjusts its orientation in the pano-rama; further evidence in accordance with this as-sumption is discussed below (3.3).

1.6.2 Input/Output AnalysisAlbeit it is hardly possible to account entirely for allstimuli any organism senses, it is probably safe toargue that all salient experience the fly receives inthe flight simulator is controlled by the experi-menter: 1) the fly is tethered and hence the visualfield of the fly is either stationary or coupled to therecording device. 2) Most other stimuli (odorants,moisture, air-pressure, magnetic or electrostaticfields etc.) are to a large degree constant during thecourse of the experiment.Consequently, most motor-output recorded is eitherinitiated by the fly on its own (initiating activity[Heisenberg, 1983] or rhythmic activity) or inducedby the experiment.

1.7 Classical and Operant Condi-tioning in DrosophilaThe aim of this study is to compare the motor-output of two groups of Drosophila fruitflies in theflight simulator which both are trained to avoid aflight direction towards a given pattern in theirenvironment. One group is trained operantly toperform the task, the other classically, according tothe definitions in 1.2 and 1.3. In such an experi-ment, it is plausible that the conditioned responsesmight deviate as an effect of the different associa-tions made during the different training procedures.More specifically, it can be expected that the oper-antly trained flies acquire ‘new’ behavioral strate-gies the classically trained flies lack or that the fliesselectively activate and inactivate certain behaviorsfrom a range of motor-programs while the classicalgroup still uses the whole range. To investigate this,the assessment techniques measuring the perform-ance of the behavior have to be identical in both theclassical and the operant experiment. The experi-mental setup has to allow for exquisite control ofstimulus presentation and response generation inorder to 1) avoid complications with stimuli unin-tentionally connected with the experiment, 2) assurethat the two experiments differ only in training and3) detect differences in response generation withsufficiently high accuracy. The flight simulator

provides the means to achieve this goal. Moreover,the concept of studying ‘microbehavior’ (Heisen-berg and Wolf, 1984) enables the student to posehis questions more specifically than in some setupsoften used by psychologists, where gross behavioris analyzed. In such preparations the expression oflearning in gross behavior is most likely to be theproduct of a rather large amount of post-acquisitionprocessing, complicating the interpretation of thedata in terms of what has been learned.

2 Material and Methods

2.1 The AnimalsDrosophila melanogaster flies of the wildtypestrain "Berlin" were used throughout the experi-ments. The flies were treated according to a breed-ing regime developed by Reinhard Wolf (pers.comm.): in order to control larval density, fliesoviposit over night on semolina pudding (AuroraHartweizen-Grieß). Using a needle, larvae and eggsare collected the next day and transferred to a vialcontaining standard cornmeal-molasses medium,keeping the larval density at 6-9 larvae/ml medium.The vials are stored in an environmental room at25°C and 60% humidity with an artificial 16hrlight/8hr dark cycle. Newly eclosed flies are trans-ferred to fresh vials on a daily basis and kept in the

Fig. 4: Illustration of typical flight orientation behavior in theflight simulator. The pattern, in this case a single vertical blackstripe, is assumed to be at infinite distance from the fly, i.e. theangular position of the stripe does not change during straightflight. a. Flight trace, consisting of torque and position trace,used to calculate the flight trajectory depicted in b. Forwardflight velocity was assumed to be constant. It can be seen thatthe torque spikes in (a) correspond with the stepwise turning ofthe panorama. (Redrawn from Heisenberg and Wolf, 1984)

6

environmental room. Vials from which flies haveeclosed for more than two days were discarded.24-48h old female flies were immobilized by cold-anesthesia and glued with head and thorax to atriangle-shaped silver wire (diameter 0.05mm) theday before the experiment (Fig. 5). The animalswere kept individually at 25°C and 60% humiditywith a 16hr light/8hr dark regime in small vials andfed a saccharose-solution until the experiment.

2.2 The Experimental SetupThe core device of the setup is the yaw torque com-pensator (Fig. 6). Originally devised by Götz (1964)and repeatedly improved by Heisenberg and Wolf(1984), it measures a fly's angular momentumaround its vertical body axis. The fly, glued to asmall hook of silver wire as described above (2.1),is attached to the torque meter via a clamp to ac-complish stationary flight in the center of a cylin-drical panorama (arena, diameter 58mm), homoge-neously illuminated from behind. Closing the feed-back loop to make the rotational speed of the arenaproportional to the fly's yaw torque (coupling factor11°/s·10-10Nm) enables the fly to stabilize the rota-tional movements of the panorama (Flight Simula-tor mode). The position of an arbitrarily chosenpoint of reference on the arena azimuth delineates aflight direction of 0-360°. Arena position (i.e. flightdirection) is recorded continuously via a circularpotentiometer and stored in the computer memory

together with yaw torque (sampling frequency20Hz).Four black, T-shaped patterns of alternating orien-tation are evenly spaced on the arena wall (widthΨ =40°, height ϑ=40°, barwidth=14°). Reinforce-ment, where applied, is made to be contiguous withthe appearance of one of the two pattern orienta-tions in the frontal quadrant of the fly's visual field.The reinforcer is a light beam (diameter 4mm at theposition of the fly), generated by a 6V, 15W Zeissmicroscope lamp, filtered by an infrared filter(Schott RG1000, 3mm thick) and focused fromabove on the fly. The strength of the reinforcer wasdetermined empirically by adjusting the voltage toattain maximum learning. In all experiments theheat was life threatening for the flies: more than 30sof continuous reinforcement were fatal for the ani-mal.The heat is applied by a computer-controlled shutterintercepting the beam (Fig. 7).

2.3 The Experiments

2.3.1 The Standard ExperimentIn this paradigm the flight simulator establishesnormal negative feedback between angular velocityof the arena and yaw torque (closed loop) through-out the experiment. This permits the animal to es-tablish optomotor balance and to adjust certainflight directions with respect to the four T-shapedpatterns on the walls of the cylindrical arena(Flight-Simulator mode). During training the fly is

Fig. 5: The flies are attached to the hook of silver wire with aUV-sensitive light in a matter of seconds. The flies are littleimpeded by the wire; they can stand, walk, and fly with it.(From: Heisenberg and Wolf, 1984)

Fig. 6: Block diagram of mode of operation of the torquecompensator for recording yaw torque of Drosophila in sta-tionary flight. HF - high frequency; LF low frequency. (redrawnfrom Götz, 1964, where a detailed description of the mode ofoperation is given)

Fig. 7: Simplified diagram of the flight simulator setup. Yawtorque is continuously transduced into d.c. voltage by the torquemeter. The computer couples this signal to the pattern drum bycalculating the angular image deviations that the measured flightmaneuvers would have caused on the fly’s eyes in free flight.Thus the fly can control the angular rotation of the drum with itsyaw torque: intended right turns of the fly rotate the arenacounterclockwise, intended left turns rotate it clockwise - similarto a human pilot controlling the panorama in a flight simulatorprogram with its joystick. The same computer controls theelectric shutter intercepting the infrared light beam used asreinforcer.

7

punished by an infrared light beam (input 6.9V),whenever one of the two pattern types is in thefrontal quadrant of the visual field. The fly cancontrol the appearance of the reinforcer by itschoice of angular pattern position. During test, theheat source is switched off.Each experiment lasts 9x2min (periods no. 1through 9). After 2x2min. of unreinforced flight(pre-test, t1), a training period of another 2x2min. isintroduced; after the following 2min. test, the2x2min. training is repeated. The experiment isconcluded by a 2x2min. test-phase (t2). At the be-ginning of each test-period the arena is rotated athigh velocity to a random angular position.A group of control-flies was subjected to the sameexperimental regime except for the reinforcement.

2.3.2 Classical ConditioningIn this paradigm, the flight simulator mode is inter-rupted (open loop) during training and the pano-rama is kept stationary with one pattern orientationin front of the fly: the fly's behavior cannot interferewith stimulus presentations. After 3s the panoramais rotated by 90o in 220ms, thus bringing the otherpattern orientation into the frontal position. One ofthe two orientations is made contiguous with thereinforcer (input 5.3V). In the test periods the appa-ratus is switched to the flight simulator mode, andthe animal's choice of angular pattern position isrecorded without reinforcement.The experimental regime is identical to the standardexperiment and consists 9 periods of 2 minuteslength (periods no. 1 through 9). After 2 periods ofunreinforced closed-loop flight (pre-test, t1), twotraining period of 2min. are introduced; after thefollowing 2min. test, the 2x2min. training is re-peated. The experiment is concluded by a 2x2min.test-phase (t2). At the beginning of each test-periodthe arena is rotated at high velocity to a randomangular position.Again the control group received the same treat-ment, but was spared the reinforcement.

2.4 The EvaluationSelf-written programs computed the raw data fromeach 2min. period into several variables of individ-ual means for each fly. The individual means wereexported into ASCII-format for later analysis in acommercial statistics program.

2.4.1 The Flight Trace I: Arena PositionAn analog to digital converter transforms the datafrom the circular potentiometer measuring the an-gular position of an arbitrary point of reference onthe cylindrical panorama (arena position) into ansequential array of consecutive data points withpossible values ranging from -2048 to 2047 (sam-pling frequency 20Hz). The zero value correspondsto an arena position at which all quadrant borders

are at a 45° angle with respect to the fly's longitudi-nal axis. These raw data are stored in the computermemory (position trace).

2.4.1.1 Avoidance/LearningAvoidance is assessed as the preference of a fly tokeep one pattern orientation in the frontal positionrather than the other. From the position trace thepreference index is calculated as (p2-pl)/(p2+pl),with p2 being the number of data points corre-sponding to a position of the arena at which thepattern orientation not associated with heat was keptin the frontal quadrant of the visual field and p1

denoting the remaining data points. A group of fliesare said to have learned if their preference index att2 (the last two periods) differs significantly fromthat of the respective control group.The mean duration of periods of staying in onequadrant (dwelling times) can also be calculatedfrom the position trace by dividing p1 and p2 by thenumber of stays in the respective sector.

2.4.1.2 FixationThe ability to keep optomotor balance with onepattern directly in front of the fly (i.e. to fly straighttowards the pattern) is assessed as the time the flykept the patterns in the frontal octant of its visualfield compared to the time the quadrant borderswere in this position. In order to calculate a measurefor fixation, the absolute values of the position tracedata array are transformed with modulo 1024 toyield values ranging from 0 to 1024. From the re-sulting array (where now the two extreme valuesrepresent the centers of two adjacent patterns) thefixation index is calculated as (f1-f2)/(f1+f2), with f2

being the number of data points n fulfilling256<n<768 and f1 being the remaining data points.

2.4.1.3 Quadrant Changes and ArenaRotationAs a measure for the activity of the fly during theexperiment, the number of quadrant changes and thetotal amount of arena rotation are calculated fromthe position trace. Adding up the events where datapoints corresponding to one pattern orientation inthe frontal position are followed by points corre-sponding to the adjacent pattern yields the numberof quadrant changes. The amount of arena rota-tion is given by the sum of the distances betweenconsecutive data points in degrees.

2.4.2 The Flight Trace II: Yaw TorqueAn analog to digital converter transforms the datafrom the torque compensator measuring the fly’syaw torque into a sequential array of consecutivedata points with possible values ranging from -2048through 2047 (sampling frequency 20Hz). The zerovalue is adjusted individually for each fly by cali-

8

brating the fly's maximum optomotor response toclockwise and counterclockwise turns of the arenato be zero-symmetrical; in the flight simulator thiscorresponds to flying straight ahead. These raw dataare stored in the computer memory (torque trace).

2.4.2.1 Spike DetectionThe spike detector used in this work takes advan-tage of the most prominent feature distinguishingspikes from the torque baseline (optomotor bal-ance): their amplitude. Since the amplitude of thetorque baseline is subjected to a considerableamount of inter- and intraindividual variation (Fig.8), detection thresholds are computed from thetorque trace every 600 data points (i.e. 30s offlight): the two peaks of a frequency distribution,gathered by arranging the data points correspondingto maxima and minima in the torque trace accordingto their frequency, delineate the interval insidewhich the torque baseline is assumed to lie (Fig. 9).

Fig. 9: Typical stretch of the yaw torque flight trace (left) with adistribution histogram of the torque maxima and minima (right).Detected spikes are indicated with arrowheads. Dotted linesdenote the detection thresholds (see text).

A continuous array of 2<n<17 data points of equalsign, the first of which exceeding the torque base-line is then considered a spike if it fulfills the fol-lowing criteria:1. Tmax>|1.4τT+5000/τT|, where Tmax denotes the

largest absolute value in the array and τT the ap-

pendant of the two thresholds delineating thetorque baseline

2. |TL|<|0.2Tmax|, or τ1<TL<τ2 where TL denotes thelast of the n data points and τ1 and τ2 denote thetwo thresholds with τ1<0<τ2.

An array of data points containing two typicalspikes is depicted in Fig. 10.

Fig. 10: Stretch of 33 yaw torque data points (i.e. 1.6s). Datapoints are connected by lines for better illustration. Dotted linesindicate detection thresholds. See text for details.

2.4.2.2 Spike Dynamics and -TimingOnce a spike is detected, its amplitude, duration andthe time elapsed since the previous spike (or sincethe quadrant change, if one occurred between twospikes) is recorded (Fig. 10). From these data threeindices are calculated: the amplitude index is cal-culated as (a1-a2)/(a1+a2) where a1 denotes the meanspike amplitude in the quadrants containing thepattern orientation associated with heat. The la-tency index is calculated as (l1-l2)/(l1+l2) with l2

being the mean time interval measured from achange into the 'hot' quadrant until the first spike inthis quadrant. The ISI index is defined accordinglyas the difference of the mean interspike intervals(ISI) in the 'hot' and the 'cold' sectors: (d1-d2)/d1+d2); here d2 denotes the mean distance be-tween two spikes in the quadrant containing thepattern orientation combined with the reinforcer.Once the timing of the spikes is accounted for, onecan calculate the number of spikes per time in eachsector and define a number index as (n1-n2)/(n1+n2)with n1 denoting the spike frequency in the ‘hot’sectors.

2.4.2.3 Spike PolarityIn addition to the force and the timing of the bodysaccades, the direction of the turns might be impor-tant for a fly when performing in the describedlearning paradigm. The polarity of a spike is de-fined as "towards pattern" if it leads to a rotation of

Fig. 8: Typical stretches of three common patterns of yawtorque flight trace. a. Oscillating mode. b. Noise mode. c.Quiet mode. (Redrawn from Heisenberg and Wolf, 1984)

9

the arena that brings the center of the nearest patterncloser to the very front, which is delineated by thelongitudinal axis of the fly. Accordingly, the spikepolarity "from pattern" brings the nearest quadrantborder closer to the most frontal position.After some preliminary calculations, the followingvariables have been taken into account: a polarityindex, yields the fraction of spikes towards thepattern. It is defined as (st-sf)/(st+sf) with st beingthe number of spikes towards the pattern and sf thenumber of spikes away from the pattern. This indexhad to be calculated for ‘hot’ and ‘cold’ sectorsseparately.The time elapsed from the entrance into a sectoruntil the first spike away from the pattern is deter-mined as well. Obviously, this is highly influencedby the timing of the spikes regardless of their polar-ity, so the difference is calculated between the la-tency index for all the first spikes towards the pat-tern and the latency index for all the first spikesaway from the pattern. If this polarity-latencyindex is positive, the first spikes away from thepattern were generated earlier in quadrants wherethe pattern orientation is associated with the rein-forcer than in the other sectors, spike frequencymodulations included.

2.4.3 The Flight Trace III: CombinedEvaluationSince torque trace and position trace are measuredsimultaneously, the torque trace can be used tocalculate the effects of torque on arena position.Angular Displacement Size. As another measureof spike dynamics the angular displacement duringeach detected spike is calculated by the distancebetween two position-data points in degrees, thecorresponding torque values of which denote theduration of the spike (Fig. 10).Stepwise Arena Rotation. As a quantitative meas-ure for the subjective impression of sudden arenarotations accounted for by spikes, the sum of angu-lar displacements during spikes is compared to thesum of interspike displacement for each 2 min.period in a rotation index: (rs-ri)/(rs+ri), where rs

denotes the sum of angular displacement during thespikes and ri the sum of arena displacements be-tween two spikes.

2.4.4 Spike Detection EfficiencyA subjective test for spike detection efficiency wasperformed with five members of the lab, only one ofwhom was familiar with Drosophila’s body sac-cades: they were presented with the task to countthe number of spikes in the torque traces of fourdeliberately chosen flies. The mean number of de-tected spikes in the first three periods was comparedto the number the spike detector counted. Two ofthe flies were deemed to be ‘regular’ flies producingspikes that were very easy to distinguish from the

torque baseline (see appendix 1 and 2). The thirdfly was flying in the ‘oscillating’ mode and thefourth fly in the ‘quiet’ mode (see appendix 3 and4).

2.5 StatisticsDuring classical training, the lack of contingencybetween behavioral output and sensory input leadsto drifting of the torque baseline over most of thetorque range of the fly. Therefore, no spike detec-tion is possible in these phases. For this reason, allcomparative studies were restricted to the five test-periods.All between-group analyses were performed withrepeated measures MANOVAs whenever more thanone 2 min period were compared at a time. Wil-coxon Matched Pairs Tests were used to test single2 min periods against zero. Correlational analyseswere computed according to Spearman’s RankOrder Correlation throughout.

3 Results and DiscussionIt was mentioned in the introduction to this studythat the different associations assumed to be madein the different training procedures - namely stimu-lus-reinforcer during classical conditioning andresponse-reinforcer during operant conditioning -might lead to different behavioral strategies to avoidthe pattern orientation associated with heat. Forinstance, if the classically trained fly learned aboutthe ‘heatedness’ of one of the pattern orientations, itmight use the same behavioral repertoire to avoidthis flight direction as is employed by the controlgroup for spontaneous preference.Conversely, operantly trained flies may have ac-quired a more effective (or at least different) way toavoid the heat during their training, selecting amongmany different behavioral strategies. In this case,the motor-output produced by those flies should bedifferent from both the respective control-group andthe classically conditioned group.It is assumed - and data supporting this assumptionis discussed below (3.3) - that torque spikes are thefly’s primary behavior to adjust flight direction(Heisenberg and Wolf, 1979; Heisenberg and Wolf,1984; Mayer et al. 1988; Heisenberg and Wolf,1993). Therefore, all properties of the torque spikesare evaluated in this study: spike amplitude, spikeduration, spike polarity, spike number, spike la-tency, interspike intervals etc.

3.1 Spike DetectionFig. 8 gives a notion how highly variable the flightmodes in different Drosophila individuals can be.As was shown by Heisenberg and Wolf (1984), sizeand shape of the torque spikes are under reafferentcontrol. As will be shown below (3.5), size, timingand polarity of the spikes are subjected to conditio-ning dependent modulations. Because of this high

10

intra- and interindividual variability in general spikeappearance, it is impossible to construct an errorproof spike detector. Even the human eye is some-times not capable of unambiguously discerningspikes from the torque baseline. In the end, only thefly ‘knows’ whether it has produced a spike or not.Fig. 11 shows this difficulty: the human subject isobviously following different rules to detect spikesaccording to the general appearance of the torquetrace. The volunteers detected less spikes than thecomputerized spike detector if the task was judgedto be ‘easy’ and spikes were clearly distinct fromthe baseline (Fig. 11, first and second fly). How-ever, when the baseline was oscillating very much(see appendix 3) the number the spike detectorproduced was lower than the number of spikes thevolunteers counted (Fig.11, third fly). The same wastrue if the baseline was very low and the spikes verysmall (see appendix 4).

Fig. 11: Comparison of the spikenumber from four flies duringthe first three periods of flight between five volunteers and thespike detector used in this study. See text for details.

Apparently, there is quite an amount of subjectivityin the detection of spikes. This is of course alsoreflected in the programmed spike detector. In thiscase the number counted by the programmer wascloser to the spike detector than the numbers of thefive volunteers (not shown). It is important, thatthroughout the study one rule is followed in order toalways compare the same turning maneuvers andthat there are not too many turning maneuvers be-tween the counted spikes which may corrupt theresults (see below, 3.3).

3.2 Spikes

Fig. 12: Mean spike numbers for all four groups (N=100 flieseach) during unreinforced flight. Open symbols indicate stan-dard groups, filled symbols indicate classical groups. Lines aredrawn for better illustration only.

The mean spike number averaged over all 400 fliesshowed a significant (p<0.001) decrease in a linearfashion from 65 spikes in the first to 29 spikes inthe last 2 min period (Fig. 12). Between-groupvariation was not significant. Heisenberg and Wolf(1979) have reported a spike frequency of 0.5-1spike/s which is confirmed by the present study.The average spike during unreinforced flight (peri-ods 1, 2, 5, 8 and 9) had an amplitude of 1.19(±0.01 S.E.M) ·10-11 Nm, which seems rather wellin accordance with the data in Heisenberg and Wolf(1979), Heisenberg and Wolf (1984) and Mayer etal. (1988). Fig. 13 shows only little variation inspike amplitude during the experiment albeit thepre-test values of the flies in the standard experi-ment are significantly higher than those of the con-trol group (p<0.001 at t1). However, the differenceis only moderate compared to the absolute values(12%) and equalized during the course of the ex-periment.

Fig. 13: Mean spike amplitudes in arbitrary units (1 unit =3.9·10-14 Nm) for all four groups (N=100 flies each) in theunreinforced periods. Note the high spike amplitude of the fliesin the standard experiment at t1 (periods 1 and 2). Open sym-bols indicate standard groups, filled symbols indicate classicalgroups. Lines are drawn for better illustration only.

11

Fig. 14: Mean spike durations during periods 1, 2, 5, 8 and 9 forall four groups (N=100 flies each). The long duration in theclassical conditioning group is obvious (see text). Lines aredrawn for better illustration only.

The mean spike duration of all 400 flies was 0.48s(±0.004s S.E.M) and in three groups it did not showmuch variation during the course of the experiment(p=0.526, Fig. 14). Only the classically conditionedflies showed a prolonged spike duration after train-ing (periods 5, 8 and 9), compared to the controlgroup (p<0.02). This is considered to be attributableto the heat, as discussed below (3.6). The measuredmean duration of a spike (0.48s) is well in accor-dance with data given in Heisenberg and Wolf(1979). Such rather long spike duration, however, isonly achieved in stationary flight without propriore-ceptive feedback from angular acceleration or reaf-ferent stimuli from air currents abruptly terminatingthe burst of torque (after 120-160ms) in free flight.The mean angular displacement caused by detectedspikes was 21.7° (±0.25° S.E.M) averaged over all400 flies. This value was rather constant throughoutthe experiment and showed little between-groupvariation (Fig. 15). This is a little less than the„roughly 30°“ reported in Heisenberg and Wolf(1979), however, it is not clear whether the numbergiven in this reference was meant to describe wild-type strain Berlin or Canton S or Drosophila ingeneral.

3.3 Stepwise Arena RotationIt was mentioned above (1.6.1) that the torque

baseline is believed to correspond to optomotorresponse behavior (optomotor balance), whereas thebody saccades (spikes) were mainly employed toadjust flight direction (Heisenberg and Wolf, 1979;Heisenberg and Wolf, 1984; Mayer et al. 1988;Heisenberg and Wolf, 1993). The rotation indexwas derived to quantify the amount of angular dis-placement accounted for by detected spikes in rela-tion to the amount of displacement between thespikes (see above, 2.4.3). The rotation index did notreveal enough between group variation in the non-reinforced periods (periods 1, 2, 5, 8 and 9) to re-ject the null hypothesis that all four groups weresamples from the same population. Therefore, thedescriptive statistics of all four groups are given inTable 1.

ConfidenceN Mean -95% +95% Median

RotInd1 400 0.35 0.33 0.38 0.39RotInd2 399 0.32 0.29 0.34 0.36RotInd8 396 0.25 0.22 0.28 0.30RotInd9 386 0.22 0.19 0.25 0.25

Min Max Lower 25% Upper 25%RotInd1 -.67 0.82 0.23 0.53RotInd2 -.70 0.79 0.17 0.51RotInd8 -.85 0.86 0.07 0.44RotInd9 -.75 0.84 0.05 0.42Table 1: Descriptive statistics of the rotation indices (RotInd) inall flies at t1 (periods 1 and 2) and at t2 (periods 8 and 9).

Considering that only between 12% and 25% of thetime of each 2min. period is consumed by spikes,the data presented in table 1 indicate that indeedmost flies rotate the arena stepwise, i.e. the flightdirection is fairly constant between the sudden turnscaused by the spikes. For instance, in the first pe-riod more than twice as much arena rotation wascaused by the spikes than by behavior in the inter-spike intervals (rotation index 0.35). The rotationindex exhibits dependence only from the meannumber of spikes per period (Spearman Rank OrderCorrelation 0.58, p<0.001 in the first and 0.50,p<0.001 in the last period). This dependence can beexemplified in two flight modes: in oscillatingmode, the spikes (if there are any) are hard to detectand much orientation might be carried out by omit-ting optomotor waggles. In quiet mode, there arevery few spikes and some orientation is accom-plished by baseline drift (personal observation).Probably the minimal values in table 1 are examplesof those flies. Of course, if a fly is producing manyspikes, there is not much room for interspike navi-gation. As the overall spike number decreases dur-ing the course of the experiment (see above, 3.2),the decrease of the rotation index is not surprising.For the same reason, some flies are not accountedfor later in the experiment: they ceased to producespikes at all.It seems that the rotation index is the lower estimateof the degree to which body saccades are used in

Fig. 15: Mean angular displacement covered by the spikes of allfour groups (N=100 each). Lines are drawn for better illustra-tion only.

12

free flight for the adjustment of flight direction: 1)as was discussed in Heisenberg and Wolf (1979)and in Heisenberg and Wolf (1984) the three flightmodes depicted in Fig. 8 may be instrumental arti-facts disturbing spike evaluation. 2) Undetectedspikes might contribute to ‘interspike’ orientation.3) Spiking behavior correlates well with learning(preference indices) in the Drosophila flight simu-lator (see below, 3.6) and thus seems indeed to beresponsible for the adjustment of flight direction.However, some flies seem to use a behavior forchoosing flight direction that is not covered by thespike detector: in about 25% of the flies at t2 (table1) the amount of summed interspike arena rotationreaches or exceeds the amount of summed arenarotation caused by spikes. Moreover, in studies ofthe Drosophila mutant strain ebo678 some fliesshowed orientation behavior without spikes, but bybaseline drift (Ilius, 1992; Ilius et al. 1994) as couldbe observed during this study with wildtype flies inthe quiet flight mode. Thus there are apparently twocomponents of orientation behavior. In the majorityof flies the dominant component is spiking behav-ior.Therefore, the evidence from the rotation index istaken as confirmation of the findings from Heisen-berg and Wolf (1979), Heisenberg and Wolf(1984), Mayer et al. (1988) and Heisenberg andWolf (1993) where torque spikes have been shownto be endogenous motor patterns (‘actions’ and notresponses to external stimuli) that are produced toadjust flight direction, whereas the baseline is as-sumed to contain the mechanism to establish andmaintain optomotor balance in a responsive wayand to be of minor importance for choosing flightdirection.

3.4 Measurements at t1

All the variables of all four groups were testedagainst the null hypothesis that they were drawnfrom the same population with a repeated measuresMANOVA for the first two periods (pre-test, t1). Inthose cases, where the null hypothesis could berejected at p<0.05 the differences were examinedmore closely. Four variables had to be taken intoconsideration (Table 2).Due to time constraints, the non-reinforced controlgroups could not be measured simultaneously withthe reinforced group (Table 3). Seasonal influences,for example, may have caused a drop in generalactivity in the non-reinforced groups. This drop isreflected in the differences in arena rotation andquadrant changes. The effect in spike amplitude ismainly due to the large spikes generated by thestandard batch, many of which were conditionedbefore the other groups. Since spike duration is

negatively correlated with spike amplitude (-0.17 atperiod 8 and -0.22 at period 9, p<0.001 in bothperiods) the result in spike duration is not surpris-ing. Taken together, the data indicate a decrease inoverall strength and activity from the reinforced tothe control groups. It has to be emphasized thatnone of the variables calculated with respect to thebehavior in the differently treated (hot/cold) sectorsshowed any deviation.

standard s-control classic c-control1st exp. 05-Oct-95 27-Feb-96 14-Jan-96 14-May-96last exp. 26-Jun-96 12-Jul.-96 02-Jul.-96 12-Jul.-96Table 3: Dates of first and last experiment in the respectivegroups.

3.5 Comparing the Standard Ex-periment and Classical ConditioningIn the standard paradigm, Drosophila avoids theheat during training very quickly and stabilizes thearena with the 'cold' pattern orientation in the fron-tal position, with short excursions into the heatedsectors. There is a very prominent behavior to beobserved during training: the entering of a rein-forced quadrant during training is often followed bya volley of spikes, bringing the fly out of the heat(Fig. 16).In the classical paradigm, Drosophila is confrontedwith the contiguity of one pattern orientation pairedwith heat during training and has no means pre-venting it from being heated. Observing flies whenheated under open loop conditions reveals a be-havior very similar to the volley of spikes depictedin Fig. 16: some flies produce spike volleys and ashift in the torque baseline during heating (Fig. 17).

standard s-control classic c-controlSpAmp*** 339.0 301.0 300.7 302.1SpDur** 0.45s 0.48s 0.48s 0.46sARot** 4742° 3646° 4206° 3691°QuCh*** 37.9 27.0 32.2 26.2Table 2: Values for those variables, where the null hypothesis thatthe four groups were drawn from one population could be re-jected, averaged over the first two periods. SpAmp - spike am-plitude (in arbitrary units; 1 unit = 3.9·10-14 Nm). SPDur - spikeduration. ARot - total amount of arena rotation. QuCh - numberof quadrant changes. * - p<0.05; ** - p<0.01; *** - p<0.001.

Fig. 16: Typical stretch of yaw torque flight trace during operanttraining with volley of spikes as the shutter opens (left arrow-head) until it closes (right arrowhead). In this trace another spikeparameter modulation can be seen: the spike amplitude is largerin the spikes during the heat than after the closure of the shutter.Dotted lines denote spike detection thresholds.

13

Fig. 17: Stretch of yaw torque flight trace of a single fly duringclassical conditioning. The pattern is switched every threeseconds. The torque trace rises high above zero together with thegeneration of large, spike-like torque fluctuations whenever theheat is switched on (arrowheads). The dotted line indicates zerotorque.

3.5.1 Avoidance and LearningIf the heat is permanently switched off, flies of bothtest groups keep their orientation preference to-wards the previously 'cold' pattern orientation for atleast several minutes (Fig. 18). In the operantgroups there was no significant difference at t1 (thefirst two periods, p=0.911), but the test groupshowed a significantly higher preference than thecontrol group (p<0.001) at t2 (the last two periods).Even if a repeated measures MANOVA revealed asignificant difference (p<0.03) in preference at t1

(the first two periods) for the classical groups, thesame analysis for all four groups (p=0.104) indi-cates that the sample flies were indeed drawn fromthe same population. Furthermore, the difference in

avoidance was of the opposite sign than that at t2

(the last two periods, p<0.027).Comparing the t2 preference indices of the operantwith the classical test group, the ‘classical’ index issignificantly lower than the operant one (p<0.015;p>0.22 for the control groups). However, when themean preference index at t1 is subtracted from theindices at t2 (to compensate for the initial individualpattern preference, see conclusion) the effect dropsbelow significance (p=0.091).In contrast to Dill et al. (1995), a comparison of themean dwelling times (periods of staying in onequadrant) for the ‘hot’ and the ‘cold’ sectors withthe respective spontaneous behavior reveals that theoperantly trained flies modulate the average timethey spend both in the ‘hot’ and in the ‘cold’ sec-tors, even during the last two periods (p<0.001 inboth cases of the operant groups, Fig. 19A). Thiscan also be seen in the mean spike amplitude in thedifferent sectors and for spike timing, respectively(data not shown). Dill et al. (1995) had found that„the dwelling times in heat associated quadrantsduring test“ were not significantly different from thecontrol group by comparing averaged individualmedians (not means as in this study) for each group.Since the frequency distributions of ‘hot’ and ‘cold’dwelling times are very similar (Reinhard Wolf,pers. comm.), the contradiction is considered not to

A

BFig. 18: Mean preference indices for all four groups of flies(N=100 each). Drosophila learns to avoid one of the patternorientations, if it was reinforced during the training periods(dotted bars). The control group, which did not receive anyreinforcement only showed random avoidance (hatched bars). A- preference indices of the standard groups, B - preferenceindices of the classical groups.

A

BFig. 19: Comparison of mean dwelling times for ‘hot’ and ‘cold’sectors (N=100 flies in each group). Individual mean dwellingtimes were averaged for each group (wide, dotted bars: testgroup, narrow, hatched bars: control group). Negative signindicates dwelling times in the reinforcer-associated quadrants.A - Standard groups, B - Classical groups.

14

be a statistical artifact. Rather the omission ofdwelling times shorter than 1s in Dill et al. (1995)might have had an influence on the significance ofthe ‘hot’ effect: they might pull the mean ‘hot’dwelling times significantly below control levels inthis study. However, since comparing in a similarmanner several of the variables discussed in detailbelow (3.5.2), produced lasting effects in the ‘hot’sectors as well (data not shown), including the shortdwelling times was probably the right choice.Classically trained flies (Fig. 19B) also change thetime they spent in one quadrant in response to thetraining, however to a lesser degree. Comparing theclassical test and control group at t2 (periods 8 and9), the differences fail to rise above the significancelevel. Only comparing the ‘hot’ dwelling times at allthree test periods yields a significant difference(p<0.04). Surprisingly, the ‘cold’ dwelling times,which subjectively seem to reveal a larger differ-ence can still not be distinguished statistically usinga repeated measures MANOVA over all three testperiods (p=0.10).

3.5.2 Spike Dynamics and -TimingWhile in neither of the groups spike duration wasmodulated in response to heat (data not shown),operantly trained Drosophila exhibited a small buthighly significant modulation of spike amplitude,even when the heat was switched off during the lasttwo periods (p<0.001, fig. 20A).Although the classically trained flies do modulatetheir spike amplitude in response to the heat as doesthe operantly trained group (Fig. 20B), the differ-ence fails to reach a level of confidence of p<0.05at t2. However, if one compares all three test-periods (nos. 5, 8 and 9) with the respective controlvalues the modulation comes to lie at a reliablep<0.03.The t2 amplitude indices of the classical test groupshow no significant deviations from those of theoperant test group (p=0.35). Furthermore, theseamplitude indices are in both groups positivelycorrelated with the respective preference indices att2 (Table 5) and before as well (data not shown).Most importantly, Drosophila seems also to modu-late the spike amplitude spontaneously: even in thecontrol groups is the amplitude index highly corre-lated with the respective preference index (Table 5).In addition to spike amplitude, the flies use thetiming of spikes to avoid the heat in the standardexperiment: as noted above, often a volley of spikesis generated to get the fly out of the heat (Fig. 21).Comparing the spike latency in the heated sectorswith that in the non-heated sectors, shows the reac-tion to the heat: the time until the first spike is sig-nificantly reduced in the ‘hot’ sectors (Fig. 21A).

This behavior is maintained even when the heat ispermanently switched off. This is not significant forthe last two periods alone, but for all three test peri-ods a p<0.041 renders the effect reliable.Since the latency indices of the operantly trainedflies were already very low at t2, it is not surprisingthat the latency indices of all three test periods fromthe classically trained batch rise only slightly abovecontrol-level (Fig. 21B). However, they could nei-ther be distinguished from the operant batch with asignificant reliability.Interestingly, both groups show different correla-tions with the preference indices in the respectiveperiods (Table 5). While there is some correlationin the operant groups - indicating that the latencyuntil the first spike has a certain predictive value foravoidance and learning in these flies - there is nocorrelation between the latency indices and thepreference of the classically trained flies at t2. Thisseems to be an effect of the open loop pattern pres-entation, since this is also revealed in the classicalcontrol group (Table 5).

A

BFig. 20: Mean amplitude indices for all four experimentalgroups (N=100 flies each). The flies learn to generate largertorque spikes in quadrants with one of the pattern orientations, ifthey were reinforced in these quadrants during the trainingperiods (dotted bars). This is the case for the operantly (A) andthe classically (B) trained flies. The control groups, which didnot receive any reinforcement modulated spike amplitude ran-domly (hatched bars).

15

Even if the timing of the first spike were not modu-lated in response to the heat, the fly still has thepossibility to make all subsequent spikes quicker toavoid the unpleasant flight direction. In the standardexperiment, the interspike intervals seem indeed tobe shorter in heated sectors than in ‘cold’ ones: fliesthat were able to control the appearance of the rein-forcer produce shorter interspike intervals in ‘hot’sectors compared to the control flies, even if theheat is switched off for the last two periods (t2,p<0.001; p=0.336 at t1; Fig. 22A).Even if the plotted ISI indices of the classical testgroup (Fig. 22B) seem rather well in accordancewith the findings in the operant test-group, they failto rise above the required significance niveau. Nev-ertheless, both pre-test values being in the negativerange and all test-values in the positive and abovethe control-values suggest a qualitatively similaralthough quantitatively less strong effect. Moreover,the corrected ISI indices at t2 (mean ISI index of thepre-tests subtracted) cannot be distinguished statis-tically from the respective values in the operantgroup (p>0.06).As the amplitude index, the distance and latencyindices of the two test groups are highly correlatedwith the preference index at t2 (Table 5). This is

also the case for the spontaneous behavior measuredin the control groups (Table 5).Having demonstrated that at least the operantlytrained flies indeed generate spikes more quickly inthose quadrants associated with the heat, one canconclude that the spike frequency (number of spikesper time) is elevated as well, which can be seen infig. 23.At first sight (Fig. 23B) the response of the classi-cally conditioned group seems to be the same as inthe operant group. However, it fails to surpass con-trol-levels to a sufficient degree (p=0.17 for allthree test periods). However, omitting period 9 withthe strikingly high control value, the modulationexceeds the control values of periods 5 and 8(p<0.02).Paralleling the conditions for the interspike inter-vals, the number indices of all four groups correlatewith the respective preference indices (see Table 5)and the corrected number indices in the classicaltest group do not differ significantly from the valuesin the operant batch (p>0.05).

A

BFig. 21: The two test groups (dotted bars) show mean latencyindices that suggest a modulation of the reaction time until thefirst spike in response to heat (see text). Most suggestive are thehigh training bars in the operant group (A). In the classicalgroup (B), only in the last period, the first spikes are producedquick enough in quadrants with the heated pattern orientationsto bring the index discernibly above the control group (hatchedbars).

A

BFig. 22: Mean ISI indices for all four groups (N=100 flies each).Drosophila produces torque spikes more quickly when it isheated in closed loop (A, dotted bars). If the heat is switchedoff, the flies still generate spikes with shorter interspike intervalsin quadrants with the ‘hot’ pattern orientation. The controlgroup, which did not receive any reinforcement only showedrandom modulation of the interspike intervals (hatched bars).In the classical groups (B), however, statistical analysis couldnot establish a significant difference between the test (dottedbars) and the control group (hatched bars). Nevertheless, thevalues are qualitatively in the proper range to assume that theyperform the same behavior as the operant group, although to alesser degree.

16

Taking together all the data presented so far, itseems as if the major reason why the effects in theclassical group appear to be less significant than inthe operant group, is the unusually low preferencein the two pre-test periods (t1) of the classicalgroup, together with the unusually high spontaneouspreference in the periods 8 and 9 (t2) of the corre-sponding control group. The finding, however, thatthe values of the classical group are not signifi-cantly different from the operant test group either istaken as evidence for the hypothesis that the effectsare indeed of the same nature as in the operantgroup. Moreover, if one expects the control groupsto exhibit symmetrical behavior, i.e. zero values inall indices, then one can test the index values ofperiod 8 with a Wilcoxon Matched Pairs Testagainst zero. This test yields significant effects forthe number index (p<0.01, period 8) and the am-plitude index (p<0.03, period 8). The ISI index isvery close to significance (p=0.067, period 8).

3.5.3 Spike PolarityIn this category, the crudest measure gave the clear-est result: in the ‘cold’ sectors, more spikes aregenerated towards the pattern than away from it,whereas in the ‘hot’ quadrants the relation is zero or

reversed. This is still true if the heat is permanentlyswitched off (Fig. 24).

A

BFig. 24: Mean polarity indices (PI) for the operantly (A) and theclassically (B) trained groups (N=100 flies each). In both groupsone can notice a rise in PI in the ‘cold’ sectors (dotted bars)compared to a rather steep decline in the ‘hot’ sectors (hatchedbars). One exception is period 8 in the classical test group.

In all the control groups, the indices for the respec-tive sectors were indiscernible and came very closeto 0.1 (Fig. 25). Comparing operant test and controlgroups at t2, both ‘cold’ and ‘hot’ indices weresignificantly different: ‘cold’ polarity indices werehigher in the test than in the control group (p<0.02)and ‘hot’ indices were lower (p<0.002). Both ‘hot’and ‘cold’ indices show significant correlations withthe respective preference indices (Table 5). As canbe expected from the low absolute differences in thepolarity indices (Fig. 24) these correlations are

A

BFig. 23: The mean number indices in the operant groups (A)clearly reflect the behavioral strategy of making many spikes inpreviously punished flight directions: whereas the control group(hatched bars) fails to produce other than random spike fre-quency modulations, the test group (dotted bars) shows directedspike frequency modulation in the predicted way.This is also visible in the classical groups (B), however to alesser degree (see text).

Fig. 25: The mean polarity indices (PI) for both control groups.Open symbols indicate the operant group, filled symbols theclassical group. Lines were drawn for illustrational purposesonly.

17

lower than those for the polarity independent vari-ables (Table 5). Furthermore, the unproportionallylarge training values commonly observed in thepolarity independent variables are missing in thepolarity indices, indicating that this parameter is notmodulated in a responsive way.As the spike polarity effects are already quite weakin the standard experiment, they diminish evenmore after classical conditioning. Evaluating thepolarity indices for the ‘hot’ and the ‘cold’ sectors(Fig. 24B) still shows a steep decline in the polarityindex for the ‘hot’ quadrants, the effect in the ‘cold’quadrants is less impressive, however. Conse-quently, only the ‘hot’ effect is statistically safe-guarded against the controls (p<0.02, taking all testperiods into account). Nevertheless, as in the stan-dard experiment, the ‘hot’ and the ‘cold’ polarityindices are significantly correlated with the prefer-ence indices (Table 5).Comparing the two test groups at t2 yields the sameresults: the ‘hot’ effect is not different in bothgroups, the ‘cold’ effect, however, is significantlylarger in the standard group (p<0.02) even if the t1

values are subtracted. This is also reflected in the

difference in fixation as will be discussed below(3.5.4).In the operant group an even more subtle but nev-ertheless significant effect can be observed in thepolarity-latency indices. (Fig. 26A). The differencebetween the test and control group is clearly to beseen throughout the experiment and at t2, when theheat is switched off, it is still statistically reliable(p<0.04). As with the polarity indices, unpropor-tionally large training values are absent, indicating agradual development of directing spike polarity.In the classical group an effect of spike polaritylatency can only be seen in period 8 (Fig. 26B). Allother periods show the same random fluctuations asthe control group.Since the absolute values are so small and the errorsare relatively large, it is not surprising that the twotest groups do not differ significantly from eachother at t2 (p=0.305).In all four groups no correlation between the polar-ity latency indices and the preference indices can bedetected (Table 5). This behavior seems thus to beacquired independently of the expression of learn-ing.

3.5.4 Variables Measured Independ-ently of Quadrant TreatmentIn both operant groups fixation increased during thecourse of the experiment, whereas in the classicalgroups only the control group showed an increase infixation (Fig. 27).

Fig. 27: Mean fixation indices of all four groups. Note that thetraining indices of the classical groups are always 1.0 due to thestatic presentation of the patterns.

Compared to the operant group, the flies that werepresented with static patterns fixated the patternssignificantly worse at t2 (periods 8 and 9; Table 4),whereas there was no difference at t1 (Fig. 27).Since the two respective control groups do notreveal any significant differences in fixation, thiseffect has to be attributed to the different rein-forcement procedures. For the same reasons, theshift in the amount of total arena rotation at t2 has tobe due to the training procedures (Table 4). Classi-cal and operant groups showed no difference in thenumber of quadrant changes (Fig. 28).

A

BFig. 26: The mean polarity-latency indices for the operant testgroups (A, dotted bars) indicate, that on average the first spikesaway from the pattern are generated earlier in quadrants wherethe pattern orientation is associated with the reinforcer than inthe other sectors (hatched bars - control group). In the classicalgroups (B), there was no detectable pattern of modulation ofspike polarity-latency neither in the test group (dotted bars), norin the control group, which did not receive any reinforcement atall (hatched bars). Only in period 8 The flies from the classicaltest group generated the first spikes away from the patternearlier in quadrants where the pattern orientation is associatedwith the reinforcer than in the other sectors.

18

Fig. 28: Combined presentation of total amount of arena rotation(upper four lines, left abscissa) and number of quadrant changesper period (lower four lines, right abscissa). Note the fixedvalues for the classical training periods.

Decreasing numbers of quadrant changes andamount of arena rotation (Fig. 28), together withdecreasing spike numbers support the view that theflies become acquainted with their situation andcalm down during the course of the experiment. Theforce of the body saccades, however, remains con-stant throughout the experiment (Fig. 13).The finding that the operantly trained flies fixatebest confirms earlier experimental data from Dill etal. (1995) where it was demonstrated that flies pre-fer flight directions in the middle of the ‘cold’ quad-rants when they have been heated in the other sec-tors. It seems as if this increase in fixation is anacquired behavioral strategy and not the conse-quence of decreasing overall activity (see below,3.6). While at t1 the fixation index is weakly (about-0.2) but significantly correlated with spike numberand total amount of arena rotation (i.e. activity) thiscorrelation ceases at t2. It is therefore not surprisingthat the flies that show the highest fixation index att2 generate more spikes and cause more arena rota-tion than the others (however, this group was al-ready more active at t1).If fixation were an actively selected behavioralstrategy, the poor fixation in the classically condi-tioned flies might be due to the avoidance of the‘hot’ pattern orientation (i.e. anti-fixation) whilefixation in the ‘cold’ sectors is within control limits.This view is discussed more thoroughly below(3.6).

standard s-control classic c-controlFixInd*** 0.52 0.41 0.26 0.37Spikes* 35.4 30.5 27.7 33.9SpDur** 0.48s 0.48s 0.54s 0.49sARot* 2986° 2470° 2285° 2658°Table 4: Values for those variables, where the null hypothesisthat the four groups were drawn from one population could berejected, averaged over the two last periods (t2). FixInd - fixationindex; Spikes - number of spikes per two minute period; SpDurmean spike duration, ARot - total amount of arena rotation. * -p<0.05; ** - p<0.01; *** - p<0.001.It cannot be ruled out, however, that the poor fixa-tion is due to fatigue caused by the high amount ofheat each fly is receiving during classical training:the flies get weaker as they are dehydrated in thecourse of the training. Consequently, they generateless and smaller spikes (Figs. 12 and 13) as well asa reduced amount of optomotor balance (quietmode, Fig. 8). Heisenberg and Wolf (1979) reportthat ‘non-fixation’ is particularly common in fliesusing the quiet mode of flight:

„[...] the stripe may suddenly start to be shifted to any positionand may be kept there quietly for some time. We call this be-havior ‘non-fixation’.“

Picturing this quiet mode might also explain theextremely long spike duration in the classical testgroup: it leads to a later reversion of the torqueslope after each spike which is the confinement ofspike duration as defined above (Fig. 10). Conse-quently, spike duration is negatively correlated witharena rotation, spike number and spike amplitude(not shown).

3.6 Measurements at t2

As discussed above (3.5.2), the flies primarilymodulate two polarity independent characteristicsof turning maneuvers in order to avoid certain flightdirections in the flight simulator: spike amplitudeand spike timing. Two sources of evidence supportthis view: the comparison of the test values with therespective spontaneous values generated by thecontrol flies (see above, 3.5.2) and a correlationalanalysis in all four groups with the preference val-ues (Table 5).

standard s-control classic c-controlAmpInd8 0.60*** 0.60*** 0.40*** 0.45***AmpInd9 0.60*** 0.68*** 0.46*** 0.46***LatInd8 0.09 0.45*** 0.15 0.14LatInd9 0.40*** 0.25* -0.01 0.14ISIInd8 0.79*** 0.68*** 0.72*** 0.83***ISIInd9 0.76*** 0.90*** 0.83*** 0.79***NumInd8 0.56*** 0.70*** 0.62*** 0.76***NumInd9 0.75*** 0.75*** 0.61*** 0.84***Table 5: Spearman Rank Order Correlations with the preferenceindex at t2. AmpInd - amplitude index, LatInd - latency index,ISIInd - ISI index, NumInd - number index. * - p<0.05; ** -p<0.01; *** - p<0.001.

19

The only obvious difference between the operantand the classical groups to be detected in Table 5 isin the latency indices. This is not surprising, sinceonly one spike is counted for each stay in one quad-rant while the fly is generating many more after-wards and therefore has many more options tochoose flight direction. Small disturbances in flightbehavior due to rather long phases of open loopmight exert a large effect in such a variable. Fur-thermore, even in those cases where the correlationis significant, the values are lower than all the othercorrelations. From the data presented so far one caninfer that flies that have high preference indicesproduce more and larger spikes in the ‘hot’ than inthe ‘cold’ sectors, even if the heat is switched off,regardless of their training procedure. Performing acorrelational analysis among the different parame-ters yields results that strengthen this view: thoseflies that generate more spikes in the ‘hot’ sectorsdo also tend make them larger (significant positiveSpearman Rank Order Correlations at t2 between theISI index, the number index and the amplitudeindex in all four groups, data not shown). This isalso the case for spontaneous behavior (see table 5):the flies use the same motor-output to express theirendogenous preference for a certain flight direction.This in turn leads to an important corollary: bothtraining procedures might modulate the endogenouspreference of each fly.Another piece of evidence pointing in this directionare some of the variables measured independentlyof quadrant quality (spike numbers, fixation indexand quadrant changes): in the test groups they aresignificantly correlated with the respective prefer-ence indices, whereas in the control groups they arenot (Table 6). Does this mean that flies that arecalmer and fixate better are also better learners?

standard s-control classic c-controlSpikes8 -0.27** -0.01 -0.27** -0.10Spikes9 -0.04 0.02 -0.14 -0.20FixInd8 0.53*** -0.10 0.34*** 0.17FixInd9 0.36*** -0.15 0.34*** 0.11QuCh8 -0.48*** 0.06 -0.47*** -0.19QuCh9 -0.23* 0.00 -0.33*** -0.16Table 6: Spearman Rank Order Correlations with the preferenceindex at t2. Spikes - number of spikes per two minute period;FixInd - fixation index; QuCh - number of quadrant changes pertwo minute period. * - p<0.05; ** - p<0.01; *** - p<0.001.

This inference is not necessarily true, of course,since performing the same analysis with the abso-lute preference indices yields identical results for allfour groups (not shown). In other words: all thereinforcer does is switching the endogenous prefer-ence in all flies to the same direction. This is mim-icked by using absolute preference indices. Then ofcourse it becomes clear why reduced activity andbetter fixation leads to higher scores: the lower theactivity is, the longer the dwelling times in eachsector and the higher the preference indices - inde-pendently of whether the fly has learned anything.

This outcome would be expected if fixation werecorrelated with the expression rather than with theacquisition of memory.Among all the variables measured independently ofthe differently treated (hot/cold) sectors, the nullhypothesis that all groups were still from the samepopulation had to be rejected in four variables:fixation index, spike number, spike duration andtotal amount of arena rotation (Table 4). Interest-ingly, the fixation indices indicate that the fliestrained in closed loop fixate the pattern moreclosely to the very front (i.e. generate more spikestowards the pattern) than do the flies presentedstationary patterns during training. As discussedabove (3.5.4), this effect is largely due to strongdeviations among the two test groups (p<0.001),whereas the control groups do not differ (p=0.449).After evaluating a number of spike polarity depend-ent variables, only the number of spikes towards thepattern compared to the number of spikes awayfrom the pattern yielded results that could be relatedto learning (Table 7).

standard s-control classic c-controlPolHot8 -0.49*** -0.30** -0.34*** -0.27**PolHot9 -0.41*** -0.50*** -0.48*** -0.43***PolCold8 0.49*** 0.42*** 0.31** 0.49***PolCold9 0.51*** 0.40*** 0.20 0.28**PolLat8 0.02 0.17 0.11 0.11PolLat9 0.05 0.19 0.02 0.14Table 7: Spearman Rank Order Correlations with the preferenceindex at t2. PolHot - ‘hot’ polarity index; PolCold - ‘cold’ polar-ity index; PolLat - polarity latency index. * - p<0.05; ** -p<0.01; *** - p<0.001.

As fixation index and overall polarity index are ofcourse highly correlated (mean Spearman RankOrder Correlation at t2 0.46, p<0.001), the polarityindices in the differently treated sectors are espe-cially telling: In the ‘cold’ sectors, more spikestowards the pattern were generated than away fromthe pattern and vice versa in the ‘hot’ quadrants.However, there was no ‘training effect’ i.e. unpro-portionally large values for the training periods,indicating that this behavior is largely independentfrom the reinforcer. This was the case for all thevariables connected with spike polarity (not shown).In the light of spike polarity, the poor fixation of theclassical test group compared to the operant testgroup might indeed reflect different behavioralstrategies acquired by the different training proce-dures: during operant training the flies learn that thecenters of the ‘cold’ quadrants are ‘safe’ (Dill et al.1995). During classical training, the sector bordersare not perceptible for the fly - the flies are pun-ished with the pattern in the centralmost position.They might even learn to avoid this central position.This experience might be more salient to the flythan the unpunished position of the other patternorientation. The data on spike polarity points in thisdirection: while in the ‘cold’ sectors the polarityindex of the classical test group does not rise above

20

the control level as does the operant group, the dropin the ‘hot’ sectors is significant and indistinguish-able from the operant group. Fig. 29 illustrates thisaccurately: in the ‘hot’ quadrants both groupsequally decrease fixation (i.e. decrease the numberof spikes towards the pattern), while in the ‘cold’sectors only the operant group decreases its staysnear the borders and increases fixation. The flies inthe classical group increased dwelling times in theentire ‘cold’ sector.This difference is also reflected in the results of acorrelational analysis among the measured behav-ioral parameters: while spike polarity at t2 in the‘hot’ quadrants was significantly correlated (nega-tively) with indices describing modulation of polar-ity independent spike parameters in all four groups,spike polarity in the ‘cold’ sectors was not corre-lated with the other indices in the classical group(Table 8).

standard s-control classic c-controlAmpIndH -0.18 -0.29** -0.27** -0.11ISIIndH -0.43*** -0.45*** -0.36** -0.22*NumIndH -0.23* -0.32** -0.30** -0.30**AmpIndC 0.23* 0.34*** 0.14 0.10ISIIndC 0.44*** 0.32** 0.21 0.43***NumIndC 0.33** 0.36*** 0.14 0.31**Table 8: Spearman Rank Order Correlations with spike polarityat period 8. Superscripts indicate with wich polarity index thevariable was correlated: H - ‘hot’ polarity index. C - ‘cold’polarity index. AmpInd - amplitude index, ISIInd - ISI index,NumInd - number index. * - p<0.05; ** - p<0.01; *** -p<0.001.

The modulation in the polarity of the first spike wasnegligible: a mean decrease of 0.06 in the probabil-ity that the first spike is towards the pattern in the‘hot’ compared to the ‘cold’ sectors was the largestvalue obtained (standard experiment). The overallprobability was over 0.8 that a first spike in anyquadrant is „towards pattern“.Summarizing the data concerning the direction ofturning maneuvers, it seems as if modulation ofspike polarity is one but not a primary behavioralstrategy for orientation in the Drosophila flightsimulator. ‘Simpler’ strategies such as modifyingthe force or the frequency of the body-saccadesseems to account for more of the avoidance behav-ior than directed orientation in space.

4 ConclusionSuppose all the assumptions made in the introduc-tion to this study are correct - namely that differentassociations are made in operant and classical con-ditioning and our paradigms represent operant andclassical conditioning - then there is only one possi-bility why the behavioral strategies of the fourgroups are so strikingly similar: the range of be-haviors a fly uses for flight orientation seems to bevery limited and rather hard wired; the componentsof this behavior are tightly interconnected. There-fore, the CR in the classical conditioning paradigmis the same as the response conditioned in the oper-ant paradigm and there is only very little room foracquiring ‘new’ strategies. An observation men-tioned above (3.5) is in favor of this notion: verysimilarly to the volley of spikes depicted in Fig. 16,some flies produce spike volleys and a shift in thetorque baseline when heated under open loop con-ditions (Fig. 17). In this view, the significant differ-ence in fixation/spike polarity might be a cue as tohow the expression of learning is accomplished inthe Drosophila flight simulator: modulating spikedynamics and -timing irrespectively of the spikes’direction seems to be a set of very basic and inter-dependent behaviors that is activated whenever thefly is asked „stay or leave?“ while the directionalusage of spikes is a more sophisticated behaviorthat becomes important when the fly is asked „howcan I stay?“ or „how can I leave?“. In the classicalconditioning paradigm studied here, the latter ques-

Fig. 29: Illustration of the changes in dwelling time distribution.From the position trace, a dwelling time/position histogram for‘hot’ and ‘cold’ quadrants is calculated for both pre-test (t1,periods 1 and 2) and test (t2, periods 8 and 9). The differencebetween these histograms at t1 and t2 is depicted here. Dottedlines indicate the centers of the patterns, vertical lines quadrantborders. The horizontal lines depict zero change in dwellingtime from t1 to t2.

21

tions are never asked, since the patterns are pre-sented stationarily during training. In the standardparadigm, however, the question how to stay awayfrom the heat is of great importance. Nevertheless,the ‘knowledge’ how to perform that task is onlyacquired slowly during the experiment - there are notraining effects in spike polarity indicating a ratherfixed, responsive behavior. Unpublished data fromReinhard Wolf (pers. comm.) lead him to the ideaof a two-process theory, too.The important implication of this hypothesis is thatfor both groups, in essence, the same has beenlearned, namely that a certain orientation of thepatterns is ‘hot’ and has to be avoided (i.e. a stimu-lus-reinforcer association has been formed). Insurplus, the flies of the operant group have learnedhow to avoid the pattern more effectively. This iswell in line with the expectations.Concentrating on the first, more basic process ofavoidance, it can be inferred from the behavior ofthe control groups that our type of conditioningmerely switches or confirms the sign of the individ-ual fly’s spontaneous preference. In addition to thedata on the flies’ behavior presented so far, it can beobserved that the preference at t1 is carried as a‘socket’ throughout the whole experiment (notshown). Then the procedure of subtracting the pre-test values (t1) from the test values (t2) is admittable.After this subtraction, only one of the indices cal-culated in dependence of the differently treated(hot/cold) sectors showed significant differencesbetween the two test groups: the ‘cold’ spike polar-ity indices (Figs. 24 and 29). All the other parame-ters showed the same modulation. If modulating thedirection of spikes is considered a more sophisti-cated strategy, maybe acquired after learning the‘basic’ avoidance task, then the flies apparentlyemployed the same behavioral output irrespectivelyof their training to perform the basic task of avoid-ing the pattern orientation associated with the rein-forcer: the flies from both groups generate many,large spikes when the previously heated patternorientation is in the frontal sector of the visual field.If the lack of significant differences in the basicresponses generated by the flies were due to thesame associations made during the different trainingprocedures, several questions have to be asked: howclassical is the standard paradigm? Is the distinctionbetween operant and classical merely operational?What is learned during conditioning?

4.1 How ‘Classical’ is the StandardExperiment?Using recorded movements of the arena (togetherwith the heating schedule) from a previously trainedfly (master) during open loop training periods in asecond fly (replay-experiments, Wolf and Heisen-berg, 1991) fails to elicit a pattern preference com-parable to that observed in the paradigms used inthis study. The recorded sequence of patterns asso-

ciated with heat sufficed to produce a significantlearning score in the master-fly, so the operantcomponent clearly is required. The classical com-ponent of pattern sequence associated with thereinforcer is not sufficient to explain learning in thestandard paradigm. However, it is not clear howthe important operant component might exert itseffect. Maybe the control of the reinforcer facili-tates a stimulus-reinforcer association; it might beeasier to perceive contingencies with one’s ownbehavior than among some stimuli totally independ-ent of oneself.If there is a ‘classical’ (learning about contingenciesin the world) association formed in the standardexperiment, the assumptions made in the beginninghave to be questioned. Either one can not consideran experiment ‘purely operant’ as soon as a singlesensory stimulus is contingent with the reinforcerand the subject learns about this stimulus. Or oper-ant conditioning is accomplished by the formationof two (or more) associations. Obviously, the viewof singular associations that are formed in the in-vestigated learning tasks was oversimplified.

4.2 Classical and Operant Condi-tioning: Merely an Operational Dis-tinction?If it was so, that whenever a single contingent sen-sory stimulus is present, the operant componentwere only facilitating the process of learning aboutthe object that is transmitting the stimulus, then alllearning in nature can only be categorized as classi-cal. There is no natural situation one can think of,where only the reinforcer is present. If this infer-ence is correct, why is ‘pure’ operant conditioning(i.e. where only the reinforcer is present; for ex-periments see e.g. Cook and Carew, 1986; Cook etal. 1991; Wolf and Heisenberg, 1991) feasible atall? So-called non-associative learning might pro-vide some useful insights, how salient situationsmight be processed in general: when rats, habituatedto a mild electric shock to their feet are transferredto a different experimental context (e.g. a larger orsmaller cage) the habituation is undetectable. Ap-parently, the rats associated the shock with theirenvironment, since transferring them back to theirtraining cage restored habituation (unindexed posteron the 24th Göttingen Neurobiology Conference,1996). This interpretation might lead to the idea ofone common mechanism, underlying all types oflearning.

4.3 What is Learned During Condi-tioning?If there is a common mechanism underlying alltypes of learning, what associations are formedduring conditioning?In the context of operant conditioning, it has to bedetermined if the introduction of a discriminative

22

stimulus into a ‘purely’ operant paradigm signifi-cantly alters the mode of acquisition of memory.One crucial experiment to find out to which degreea classical component is integrated in an operantparadigm, seems to be to train an individual oper-antly to use one output channel to optimize itsstimulus situation (reinforcer and stimulus) and thentest it by coupling a different output channel to thesame environment (stimulus without reinforcer).Such an experiment has not been performed, yet.In the context of classical conditioning, the issue iswhich properties of the contiguous events becomeencoded and associated. One consequence of thenotion that stimulus-reinforcer associations wereformed during classical training has gathered muchattention in the literature: if simple stimulus substi-tution were to account for learning, the CR shouldbe identical to the UR. Several observations seem toindicate that this is not the case. Rabbits for in-stance, respond with swallowing and jaw move-ments during training in a salivary conditioningparadigm, but fail to show these behaviors duringtest (Sheffield, 1965; cited in Rescorla and Holland,1982). Or the CR might include behaviors not pres-ent in the UR: Pavlov’s abovementioned dog thatshowed appetitive behavior towards the bell is oneexample of motor activity towards CSs paired withfood, although activity is not part of the response tofood itself. Pigeons peck visual signals for USs thatdo not elicit pecking, such as water delivered di-rectly into the mouth (Woodruff and Williams,1976; cited in Rescorla and Holland, 1982) or heat(Wasserman, 1973; Wasserman et al. 1975; cited inRescorla and Holland, 1982). Spear et al. (1990)cite Pinel et al. (1980) where conditioning is ex-pressed by suppression if a tone predicted the US(shock) but by active hiding if the US was signaledby a prod.As noted above (4), some flies observed during thisstudy confirm the evidence in the literature: theyproduced spike volleys and a shift in the torquebaseline when heated under open loop conditions(Fig. 17), clearly to be classified as an UR to theheat. This behavior disappears completely if theheat is switched off (Reinhard Wolf, pers. comm.);a trace of it, however, can be detected in closedloop: the spikes produced in quadrants with thepreviously heated pattern are larger and closer to-gether than in the other sectors.Moreover, it has been shown that conditioning stilldoes occur if the stimulus-properties of the US aresuppressed: stimulation of the VUMmx1-neuron inbees can serve as substitution for the sugar-reinforcer (Hammer, 1993). Suppression of theresponse-evoking properties of the US, for instanceby applying response attenuating drugs such ascurare does not prohibit learning either (Solomonand Turner, 1962; cited in Rescorla and Holland,1982), ruling out direct stimulus-response associa-tions in classical conditioning.

If the view of singular response-reinforcer orstimulus-reinforcer or stimulus-response associa-tions is so oversimplified, what is happening in thebrain of a conditioned organism?Assume an animal struggling for survival: everysensation might provide a clue how to escape apredator, find a mate, explore new food patches,hiding places, etc. In every second it is confrontedwith potentially dangerous or advantageous situa-tions. The possibility to predict such situations mustconvey an enormous selection pressure. A veryeffective way to accomplish this task would be anevaluation mechanism, judging situations accordingto their ‘beneficence’ for the individual. With sucha mechanism salient internal and external stimulus-arrays extracted from the situation would receivesituation-specific rankings on a value-scale in termsof ‘good’, ‘bad’ or ‘neutral’. The probability ofperforming a given behavior in a certain stimulussituation is the manifestation of a more or less com-plex superposition of stimulus-rankings, motivationand initiating activity. In this picture learning corre-sponds to linking neutral or unknown stimuli toalready ranked ones, whenever a sufficient cross-correlation between them is detected. In ‘pure’operant conditioning, the internal representation ofbehaviors (efference copy, von Holst and Mit-telstaedt, 1950) is linked to the ranking of the rein-forcer if the correlation coefficient between them issufficiently positive. If more stimuli are contingentwith the reinforcer, they receive adequate rankingsas well. The richer the environment, the more com-plex the net between the stimuli becomes. An apriori ranking of stimuli and behavioral representa-tions constitutes the basis for URs, fixed actionpatterns, and the species-specific salience and asso-ciability of certain stimuli. As mentioned above,many of these rankings are assumed to be situationspecific. Situation-specific in this case means thesituation in which the ranking has been acquired.For instance, if the reinforcer was food, the con-tiguous behavior(s) or event(s) would receive afood-specific ranking rather than a mate- or danger-specific ranking. Prolonged training intensifies thelinks (associations) between the stimuli. This mightlead to such prominent behavior as described abovefor the chimpanzees.Such an informal model is in conformity with sev-eral attentional models for acquisition (e.g. Rescorlaand Wagner, 1972; Mackintosh, 1975), describingovershadowing (Pavlov, 1927) or blocking (Kamin,1969), since the distance of stimuli from the value‘neutral’ might convey them with the appropriateattentional properties. It fully accounts for differ-ences in UR and CR by not only transferring the USvalue to the stimulus, but also by linking the rank-ings of different US features to the appropriatebehavioral rankings (situation-specificity). Sensorypreconditioning (in rats: Rescorla and Cunningham,1978; in bees: Müller et al. 1996) is predicted by

23

this model as the ranks of contiguous (or similar forthat matter) stimuli would become linked.Such a model would imply that as soon as one sali-ent sensory stimulus is presented contiguous withthe reinforcer in an operant conditioning paradigm,the subject will use all output-channels to respondappropriately during test.

5 SummaryThe assumptions for this study were as follows: inoperant conditioning a response-reinforcer associa-tion is formed, whereas in classical conditioning theassociation is stimulus-response in nature. Using theDrosophila flight simulator, the motor-output oftwo groups of Drosophila flies - one presented withstationary stimulus-reinforcer pairings (open looptraining), one enabled to control the presentation ofthe pairings (closed loop training) - was compared.It was assumed that the open loop group was classi-cally, the closed loop group operantly conditioned.If this were correct, there might be a difference inthe behavioral strategies an operantly trained sub-ject might acquire compared to a subject classicallytrained to perform the same task. It can be con-cluded from the gathered data that one of the ob-served behavioral strategies was significantly en-hanced in the operantly but not in the classicallytrained flies: choosing a flight direction as far awayfrom the reinforced stimulus as possible.It appears, however, that an individual during con-ditioning is not confined to making singular asso-ciations but is rather evaluating complex stimulussituations. This leads to the transfer of reinforcerproperties to the unconditioned stimulus even in theoperant paradigm. An experiment in which thesubject uses one behavioral repertoire (e.g. walkingbehavior) to control the reinforcer and a sensorystimulus during training and another (e.g. flightcontrol behavior) to control the conditioned stimu-lus during test, deserves special interest in resolvingthis issue.

6 AcknowledgementsMost of all I should like to thank my parents forproviding me the possibilities to endeavor intoscience. Next, of course, my gratitude goes to Prof.Dr. Martin Heisenberg for his constant encourage-ment during this study and for the enlighteningdiscussions without which the progress of this workwould have been at stake. Of inestimable value formy study was Reinhard Wolf without whom all theprograms and other equipment would not haveworked. Moreover, I profited much by his experi-ence, his intensive feedback and constant flow ofideas. I also want to thank Matthias Porsch, MartinBarth, Reinhard Wolf, Jean-René Martin andGerold Wustmann for volunteering for the test ofspike detection efficiency. I’m also indebted toChris Cutts from the University of Glasgow for

checking the thesis for language errors. During allof my time working I have been inspired by inter-esting discussions during our weekly sessions in thebehavior group and by the congenial atmosphere inthe Department of Genetics and the Biocenter.

7 ZusammenfassungUnter der Annahme, daß ein Subjekt bei operantemKonditionieren eine Kreuzkorrelation zwischeneigenem Verhalten und seiner Stimulus-Situationbildet und bei hinreichend positiver Übereinstim-mung eine Antwort-Verstärker Assoziation knüpft,bei klassischem Konditionieren jedoch Eigenschaf-ten des konditionierten Stimulus durch Eigenschaf-ten des Verstärkers ersetzt werden (Stimulus Sub-stitution) und somit eine Stimulus-Verstärker Asso-ziation vorliegt, wurde in dieser Studie das Verhal-ten von zwei Gruppen von Drosophila Fruchtflie-gen am Flugsimulator verglichen. Den Fliegen dereinen Gruppe wurden zwei stationäre Musterpaarezusammen mit einem Verstärker in „open loop“präsentiert, die Fliegen der anderen Gruppe konntendie Muster- zusammen mit der Verstärkerpräsenta-tion selbst kontrollieren (closed loop). Es wurdedavon ausgegangen, daß erstere Gruppe klassisch,letztere operant konditioniert worden sei.Bei Richtigkeit aller Annahmen war eine Selektionvon ‘effektiven’ Strategien im Vermeideverhaltender operant trainierten Gruppe zu erwarten. Aus dengewonnenen Daten kann geschlossen werden, daßeine der beobachteten Verhaltensstrategien nur vonden operant trainierten Fliegen selektiv aktiviertwird und nicht von den klassisch trainierten, näm-lich eine Flugrichtung mit größtmöglicher Entfer-nung vom bestraften Musterpaar einzuschlagen.Die Annahme singulärer Assoziationsbildung er-scheint jedoch übersimplifiziert. Offensichtlichwerden alle mit dem Verstärker in Zusammenhangzu bringenden Sachverhalte ausgewertet. Dies führtauch im ‘operanten’ Paradigma zu einem Transfervon Verstärkereigenschaften auf den konditionier-ten Stimulus. Von besonderem Interesse zur Klä-rung dieses Sachverhaltes wäre ein Experiment, indem im Training ein Verhaltensrepertoire (z.B.Laufverhalten) den Verstärker und einen sensori-schen Stimulus kontrolliert und im Test ein anderesRepertoire (z.B. Flugverhalten) zur Kontrolle deskonditionierten Stimulus eingesetzt werden muß.

8 ReferencesCarew T, Hawkins R, Kandel E (1983): "Diffe-

rential classical conditioning of a defensivewithdrawal reflex in Aplysia californica."Science. 219: 397-400

Carew T, Sahley C (1986): "Invertebrate learningand memory: from behavior to molecules."Annu Rev Neurosci. 9: 435-87

24

Cook D, Carew T (1986): "Operant conditioningof head waving in Aplysia." Proc Natl AcadSci U S A. 83: 1120-4

Cook D, Stopfer M, Carew T (1991): "Identifica-tion of a reinforcement pathway necessary foroperant conditioning of head waving in Aply-sia californica." Behav Neural Biol. 55: 313-37

Davis R, Dauwalder B (1991): "The Drosophiladunce locus: learning and memory genes inthe fly." Trends Genet. 7: 224-9

Dill M, Wolf R, Heisenberg M (1995): "Behavio-ral analysis of Drosophila landmark learningin the flight simulator." Learning and Memory2: 152-60

Eyding D (1993): "Lernen und Kurzzeitgedächtnisbeim operanten Konditionieren auf visuelleMuster bei structurellen und bei chemischenLernmutanten von Drosophila melanogaster."Diplomarbeit, Würzburg, pp. 85

Glanzman D (1995): "The cellular basis of classi-cal conditioning in Aplysia californica - it'sless simple than you think." Trends Neurosci.18: 30-6

Götz K (1964): "Optomotorische Untersuchungdes visuellen Systems einiger Augenmutantender Fruchtfliege Drosophila." Kybernetik. 2:77-92

Götz K (1983): 'Bewegungssehen und Flugsteue-rung bei der Fliege Drosophila.' In: Nachti-gall, W. (ed.): Biona Report 2 Fischer, Stutt-gart, pp. 21-34

Götz K, Hengstenberg B, Biesinger R (1979):"Optomotor control of wingbeat and body po-sture in Drosophila." Biol Cybern. 35:101112

Hammer M (1993): "An identified neuron media-tes the unconditioned stimulus in associativelearning in honeybees." Nature. 366: 59-63

Hawkins R, Abrams T, Carew T, Kandel E(1983): "A cellular mechanism of classicalconditioning in Aplysia: activity-dependentamplification of presynaptic facilitation."Science. 219: 400-5

Heisenberg M (1983): "Initiale Aktivitat und Will-kürverhalten bei Tieren." Naturwissenschaf-ten. 70: 70-8

Heisenberg M, Wolf R (1979): "On the finestructure of yaw torque in visual flight orien-tation of drosophila melanogaster." J compPhysiol. 130: 113-30

Heisenberg M, Wolf R (1984): "Vision in Dro-sophila. Genetics of Microbehavior." Sprin-ger, Berlin, Heidelberg, New York, Tokio, pp.250

Heisenberg M, Wolf R (1993): "The sensory-motor link in motion-dependent flight controlof flies." Rev Oculomot Res. 5: 265-83

Holst E von, Mittelstaedt H (1950): "Das Reaffe-renzprinzip. Wechselwirkungen zwischen

Zentralnervensystem und Pripherie." Natur-wissenschaften 37: 464-76

Ilius M (1992): "Untersuchungen der visuellenFlugsteuerung bei Mutanten des Gens el-lipsoid-body-open (ebo) von Drosophila me-lanogaster." Diplomarbeit, Würzburg, pp. 112

Ilius M, Wolf R, Heisenberg M (1994): "Thecentral complex of Drosophila is involved inflight control: studies on mutants and mosaicsof the gene ellipsoid body open." J Neuroge-net. 9: 189-206

Kamin L (1969): 'Predictability, surprise, attenti-on, and conditioning.' In: Campbell, B., andChurch, R. (eds.): Punishment and aversivebehavior Appleton-Century-Crofts, New Y-ork, pp. 279-96

Kandel E, Abrams T, Bernier L, Carew T, Haw-kins R, Schwartz J (1983): "Classical con-ditioning and sensitization share aspects ofthe same molecular cascade in Aplysia." ColdSpring Harb Symp Quant Biol. 48: 821-30

Locke J (1689): "An essay concerning humanunderstanding." ILT Digital Classics,http://www.ilt.columbia.edu/academic/digitexts/locke/understanding/title.html

Mackintosh N (1975): "A theory of attention:Variations in the associability of stimuli withreinforcement." Psychol Rev. 82: 276-98

Mayer M, Vogtmannn K, Bausenwein B, WolfR, Heisenberg M (1988): "Flight control du-ring 'free yaw turns' in Drosophila." J CompPhysiol. 163: 389-99

Müller D, Gerber B, Hellstern F, Hammer M,Menzel R (1996): 'Within-compound associ-ations in honeybees.' In: Elsner, N., andSchnitzler, H.U. (eds.): Proceedings of the24th Göttingen Neurobiology Conference1996 Thieme, Stuttgart, New York, p. 769

Pavlov I (1927): "Conditioned reflexes." OxfordUniversity Press, Oxford

Pinel J, Treit D, Wilkie D (1980): "Stimuluscontrol of defensive burying in the rat." LearnMotiv. 11: 150-63

Rescorla R, Cunningham C (1978): "Within-compound flavor associations." J Exp PsycholAnim Behav Process. 4: 267-75

Rescorla R, Holland P (1982): "Behavioral stu-dies of associative learning in animals." AnnuRev Psychol. 33: 265-308

Rescorla R, Wagner A (1972): 'A theory of Pavlo-vian conditioning: Variations in the effective-ness of reinforcement and nonreinforcement.'In: Black A.H.; Prokasy W.F. (eds.): ClassicalConditioning II: Current research and theoryAppleton-Century-Crofts, New York, pp. 64-99

Sheffield F (1965): 'Relation of classical conditio-ning and instrumental learning.' In: Prokasy,W. (ed.): Classical Conditioning Appleton-Century-Crofts, New York, pp. 302-22

25

Skinner B (1938): "The behavior of organisms."Appleton, New York

Solomon R, Turner L (1962): "Discriminativeclassical conditioning in dogs paralyzed bycurare can later control discriminative avoi-dance responses in the normal state." PsycholRev. 69: 202-19

Spear N, Miller J, Jagielo J (1990): "Animalmemory and learning." Annu Rev Psychol.41: 196-211

Tully T (1991): "Of mutations affecting learningand memory in Drosophila-the missing linkbetween gene product and behavior." TrendsNeurosci. 14: 163-4

Tully T, Boynton S, Brandes C, Dura J, MihalekR, Preat T, Villella A (1990): "Genetic dis-section of memory formation in Drosophila ."Cold Spring Harb Symp Quant Biol. 55: 203-11

Tully T, Preat T, Boynton S, Del-Vecchio M(1994): "Genetic dissection of consolidatedmemory in Drosophila." Cell. 79: 35-47

Wasserman E (1973): "The effect of redundantcontextual stimuli on autoshaping the pige-

on's key peck." Anim Learn Behav. 1: 198-201

Wasserman E, Hunter N, Gutowski K, Bader S(1975): "Autoshaping chicks with heat rein-forcement: The role of stimulus-reinforcerand response-reinforcer relations." J ExpPsychol.: Anim Behav Proc. 1: 158-69

Weidtmann N (1993): "Visuelle Flugsteuerungund Verhaltensplastizität bei Zentralkomplex -Mutanten von Drosophila melanogaster."Diplomarbeit, Würzburg, pp. 87

Wolf R, Heisenberg M (1991): "Basic organizati-on of operant behavior as revealed in Dro-sophila flight orientation." J Comp Physiol A.169: 699-705

Woodruff G, Williams D (1976): "The associativerelation underlying autoshaping in the pige-on." J Exp Anal Behav 26: 1-14

26

Appendix 1: Flight trace of fly no. 1 that was evaluated by five volunteers for spike numbers. Thick curve: posi-tion trace, thin curve: torque trace. Shaded areas: ‘hot’ sectors. Each block contains a two minute period.

27

Appendix 2: Flight trace of fly no. 2 that was evaluated by five volunteers for spike numbers. Thick curve: posi-tion trace, thin curve: torque trace. Shaded areas: ‘hot’ sectors. Each block contains a two minute period.

28

Appendix 3: Flight trace of fly no. 3 that was evaluated by five volunteers for spike numbers. Thick curve: posi-tion trace, thin curve: torque trace. Shaded areas: ‘hot’ sectors. Each block contains a two minute period.

29

Appendix 4: Flight trace of fly no. 4 that was evaluated by five volunteers for spike numbers. Thick curve: posi-tion trace, thin curve: torque trace. Shaded areas: ‘hot’ sectors. Each block contains a two minute period.

30

Ich erkläre hiermit, daß ich die vorliegende Arbeit selbständig verfaßt und keine anderen als die angegebenenQuellen und Hilfsmittel verwendet habe.

Würzburg, den 30.08.1996